US11502879B2

US11502879B2 - Multi-stage equalizer for inter-symbol interference cancellation

Info

Publication number: US11502879B2
Application number: US17/265,294
Authority: US
Inventors: Prashant Choudhary; Nanyang Wang
Original assignee: Rambus Inc
Current assignee: Cadence Design Systems Inc
Priority date: 2018-08-17
Filing date: 2019-07-31
Publication date: 2022-11-15
Also published as: EP3837817A1; US20230119007A1; EP3837817A4; US11876650B2; CN112567703A; US20210306187A1; WO2020036740A1

Abstract

An equalizer includes a first feed-forward stage that provides a measure of low-frequency IS I and a second feed-forward stage that includes a cascade of stages each making an IS I estimate. The IS I estimate from each stage is further equalized by application of the measures of low-frequency IS I from the first feed-forward stage and fed to the next in the cascade of stages. The IS I estimate from the stages thus become progressively more accurate. The number of stages applied to a given signal can be optimized to achieve a suitably low bit-error rate. Power is saved by disabling stages which are not required to meet that goal.

Description

BACKGROUND

Binary communication systems represent information using just two symbols—e.g. relatively high and low voltages—to alternatively represent a logical one and a logical zero (i.e. 1 or 0). The number of levels used to represent digital data is not limited to two, however. For example, a type of signaling referred to as PAM-4 (for 4-level pulse-amplitude modulation) provides for four discrete pulse amplitudes (voltages) to convey two binary bits of data per symbol (i.e., 00, 01, 10, or 11). A series of symbols, each representing one of a set of possible symbol values, can thus be communicated as a voltage signal that transitions between levels in a manner that reflects the series. The time each voltage level is held to represent a symbol is termed the “symbol time” and the speed with which symbols can be communicated is termed the “symbol rate.” A receiver recovers a set of symbols from a signal by comparing the voltage during each symbol time against one or more reference-voltage levels to distinguish between symbols.

High performance communication channels suffer from many effects that degrade symbols, and consequently render them difficult to resolve. Primary among them are frequency dependent channel loss (dispersion) and reflections from impedance discontinuities. These effects cause neighboring symbols (voltage levels) to interfere with one another and are commonly referred to collectively as Inter-Symbol Interference (ISI). For example, relatively high-voltage symbols may spread out to raise the level of neighboring lower-voltage symbols; if the resulting voltage distortion is sufficiently high, the lower-voltage symbols may be interpreted incorrectly. Lower-voltage symbols may likewise induce errors in neighboring higher-voltage symbols.

ISI becomes more pronounced at higher signaling rates, ultimately degrading signal quality such that distinctions between originally transmitted symbols may be lost. Some receivers mitigate the effects of ISI using one or more equalizers to reverse the effects of channel distortion. One common type of equalizer, the decision-feedback equalizer (DFE), corrects for ISI imposed on a current symbol by a prior symbol by multiplying the value of the prior symbol by a “tap value” calibrated to account for the ISI and adding the resultant product to the voltage level used to interpret the current symbol. ISI from a prior symbol that tends to raise (lower) the level of the current symbol is thus offset by a similar rise (fall) in the reference against which the current symbol is evaluated. The same principle can be extended to multiple preceding symbols.

In high-speed systems it can be difficult to resolve recently received symbols in time to calculate their impact on incoming symbols and apply the requisite feedback. Symbol pre-decision is used in some DFEs to ease this timing constraint. Each received symbol is ISI-compensated based on hypothetical decision feedback signals and sampled against decision reference levels to make tentative symbol pre-decisions. The decision feedback loop is then used to select among the tentative pre-decisions after the prior symbol is resolved.

The use of pre-decisions to delay the need for a final decision on a prior symbol reduces the latency of the decision-feedback loop and helps timing closure of the timing critical path. Unfortunately, this advantage requires increasing the requisite number of samples for each symbol, and thus the required circuit area and power. Binary signaling requires two tentative decisions for each prior symbol under consideration. PAM-4 signaling requires twelve tentative decisions, which results in significantly higher power usage and circuit area relative to binary signaling and makes timing closure more difficult.

BRIEF DESCRIPTION OF THE DRAWINGS

The detailed description is illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings and in which like reference numerals refer to similar elements and in which:

FIG. 1 depicts a receiver 100 that combines ISI estimates with feed-forward equalization to accurately and efficiently digitize an analog, continuous-time input signal Vin(t) distorted by inter-symbol interference (ISI).

FIG. 2 depicts a receiver 200 that combines ISI estimates with feed-forward equalization to accurately and efficiently digitize an analog, continuous-time input signal Vin(t) distorted by ISI.

FIG. 3 depicts a PAM4 receiver 300 that receives an analog, continuous-time input signal Vin(t) at a rate 56 GBaud.

FIG. 4 depicts a waveform diagram 400—a so-called “eye diagram”—of a PAM4 signal input to a decision circuit 355.

FIG. 5 depicts ISI estimating Stage 1 of FIG. 3 with decision threshold and equalizer tap adaptation circuitry 500.

DETAILED DESCRIPTION

FIG. 1 depicts a receiver 100 that combines estimates of inter-symbol interference (ISI) with feed-forward equalization to accurately and efficiently digitize an analog, continuous-time input signal Vin(t) distorted by the ISI. An analog-to-digital converter (ADC) 105 converts signal Vin(t) into a digital input signal Din. A feed-forward equalizer FFE1 equalizes input signal Din to produce an equalized signal EQ1 that provides a measure of high-frequency components of the ISI. Delay elements 110 provide delayed versions of equalized signal EQ1 each synchronized to one stage of a cascaded ISI canceller ISIC. Each stage of ISI canceller ISIC provides high-confidence and low-confidence ISI estimates expressing tentative values of input signal Din, equalizes these tentative values to produce ISI estimates, and combines the ISI estimates with a time-aligned version of equalized signal EQ1 to provide an improved representation of input signal Vin(t). A multiplexer 120 can take a hard-decision output HDQ, a digitized version of input signal Vin(t), from any of three decision circuits 115. One or both stages of ISI canceller ISIC can thus be disabled to save power if not required to establish a desired bit-error rate (BER). Receiver 100 can thus be power optimized for channels of varying quality.

Feed-forward equalizer FFE1 equalizes signal Din responsive to a first set of tap values (not shown). In this context, a “tap value” is a coefficient applied to a symbol propagating through a delay line. The current symbol under consideration is referred to as the “main cursor” and the corresponding tap value the “main tap.” Symbols before or after the main cursor are respectively termed pre-tap and post-tap and may be designated by the number of symbol times they are separated from the main cursor. Given a main symbol S(n) with tap T(n), for example, the prior symbol may be designated S(n−1) and the prior tap T(n−1). Filters and filter taps are well understood by those of skill in the art, so a detailed discussion is omitted.

The leftmost decision circuit 115 produces a binary hard decision HDQ1 for each symbol of equalized signal EQ1. In this context, a “hard decision” is a high-confidence resolution of a symbol from among an expected set of symbol values expressed by the sampled signal. In this binary example, the expected set of symbol values consists of binary one and zero. Decision circuit 115 can thus resolve a given symbol as either a binary one or a binary zero. For short channels filter FFE1 may provide enough equalization that merely sampling signal EQ1 will produce a sufficiently low BER. In that case multiplexer 120 can select signal HDQ1 as receive data HDQ and the unused elements of receiver 100 can be disabled to reduce power consumption.

ISI canceller ISIC includes a cascade of stages, two stages ISIC1 and ISIC2 in this embodiment. Considering first stage ISIC1, a decision circuit 125 slices equalized signal EQ1 relative to two reference levels TH1 ₁:TH1 ₀to obtain high-confidence decisions and low-confidence decisions. Decisions of low and high confidence resolve a symbol from among a set of symbol values greater in number than what is expressed by the sampled signal. In this example in which equalized signal EQ1 represents binary data, decision circuit 125 designates slices indicative of three values, those with levels greater than TH1 ₁as high-confidence logic-one decisions, less than TH1 ₀as high-confidence logic-zero decisions, and levels between TH1 ₁and TH1 ₀as low-confidence indeterminant-X decisions that might represent either a logic one or a logic zero.

FIG. 1 includes a waveform diagram 150—a so-called “eye diagram”—of binary equalized signal EQ1 input to decision circuit 125. Diagram 150 is used to illustrate the distinction between high-confidence regions HC1 and HC0 and low-confidence region LC. Signal EQ1 includes an “eye” 155 separating the two voltage ranges that represent the two possible binary values at a sample instant around time zero. Sample values of signal EQ1 that fall within the range encompassing eye 155 do not clearly resolve the symbol under consideration. Decision circuit 125 slices incoming symbols relative to two threshold levels TH1 ₁:TH1 ₀to place each symbol into one of the three regions. Decision circuit 125 outputs soft-decision signal SDQ1 placing each sliced symbol from signal EQ1 into one of the high- or low-confidence regions.

Confidences specified as “high” or “low” represent relative values, with decision circuit 125 producing outputs representative of just two levels in this example. Decision circuits in accordance with other embodiments can produce outputs representative of more and different confidence levels in other embodiments. The voltage range of a decision circuit might be spanned by e.g. five or seven regions of confidence, the region of lowest confidence bounding zero volts and regions of progressively higher confidence toward the high and low voltage extremes. One such embodiment is discussed below in connection with FIG. 4.

An equalizer 130 equalizes soft-decision signal SDQ1 to produce a second equalized signal EQ2, a second equalized version of the input signal and estimate of ISI. Equalizer 130 applies full tap values to high confidence decisions to offset interference from aggressor symbols. Equalizer 130 applies scaled-down tap values (e.g. no offset) to low-confidence decisions to limit error propagation. An arithmetic operator 135 combines the resultant ISI estimate IE1 with a time-aligned version of equalized signal EQ1 to produce a second equalized signal EQ2. The delay imposed by delay element 110 on signal EQ1 matches the symbol timing within stage ISIC1.

In this example, equalizer stage FFE1 cancels relatively low-frequency ISI, which is to say that stage FFE1 cancels the tails of the single-bit response of the channel by which signal Vin(t) arrives. The output of stage FFE1, and similarly delay element 110, is thus a measure of higher-frequency ISI that can be employed to reduce the BER of stage ISIC1. If equalized signal EQ2 from stage ISIC1 produces a sufficiently low BER then multiplexer 120 can be controlled to select the output HDQ2 from a second of decision circuits 115 as receive data HDQ. As before, any unused elements of receiver 100 can be disabled. Receiver 100 can thus be power optimized for channels of varying quality.

The output from stage ISIC1 may have an undesirably high BER, such as for relatively long channels. In that case the second stage ISIC2 can be engaged to further reduce the BER. Stages ISIC1 and ISIC2 are similar in this embodiment. The stages differ, however, in the levels used to distinguish high-confidence decisions from low-confidence decisions. Each successive stage of equalizer ISIC receives an equalized signal with a lower BER than the one preceding. The levels employed to distinguish decisions reflect the resultant changes in decision boundaries. Decision circuit 125 in stage ISIC2, for example, slices equalized signal EQ2 to place them relative to reference levels TH2 ₁:TH2 ₀to produce a second series of high-confidence and low-confidence decisions SDQ2. The low-confidence central region LC for stage ISIC2 spans a different range of amplitudes than region LC from preceding stage ISIC1. Second stage ISIC2 produces another ISI estimate IE2, which is combined with yet another delayed version of equalized signal EQ1 to produce a third equalized signal EQ3. Multiplexer 120 can select the output HDQ3 from a third of decision circuits 115 as receive data HDQ.

Decision circuit

125 is a soft-decision slicer in this example but can be e.g. a hard-decision slicer, a most-significant-bit (MSB) slicer, a decision-feedback equalizer (DFE), a maximum-likelihood sequence detector (MLSD), or a maximum a posteriori probability (MAP) detector. Moreover, different ISI estimating stages ISIC# can use different forms of decision circuits and equalizers, and one or more additional ISI estimators with the same or fewer stages than ISI canceller ISIC can be included in parallel with the one shown. Each decision circuit 115 is a hard slicer in this embodiment but can otherwise be e.g. a DFE, an MLSD, or a MAP detector.

FIG. 2 depicts a receiver 200 that combines ISI estimates with feed-forward equalization to accurately and efficiently digitize an analog, continuous-time input signal Vin(t) distorted by ISI. Receiver 200 is much like receiver 100 of FIG. 1, with like-identified elements being the same or similar. Receiver 200 differs from receiver 100 in that the input of ISI canceller ISIC comes from a second feed-forward equalizer FFE2 that delivers a second equalized version of the input signal EQ2′. Cascaded ISI-estimation stages ISIC1 and ISIC2 work as detailed previously to produce decisions SDQ1′ and SDQ2′, ISI estimates IE1′ and IE2′, and equalized signals EQ3′ and EQ4. Multiplexer 120 selects from hard decisions HDQ2′, HDQ3′, and HDQ4 depending upon the level of equalization required to obtain a desired symbol rate and BER. In other embodiments, ISI canceller ISIC can produce ISI estimates from equalized signal EQ2′ and one or more additional equalized signals (e.g. equalized signal EQ1).

FIG. 3 depicts a PAM4 receiver 300 that receives an analog, continuous-time input signal Vin(t) at a rate of 56 GB aud. Each symbol represents one of four pulse amplitudes to convey two binary bits of data per symbol for a data rate of 112 Gbs. An ADC 305 converts signal Vin(t) into a series of six-bit digital values Din(n) that cover a range of voltages from between about 0.3V to −0.3V in sixty-four steps (e.g., from decimal −32 to +31). The port and signal path for signal Din are labeled in a fixed-point numbering format identified by a leading “Q” in which integers preceding and following a decimal identify numbers of integer and fractional bits, respectively, and the sign bit is excluded. The notation Q5.0 thus represents six bits; an implied sign bit, five integer bits, and zero fractional bits.

Receiver

300 includes P slices 310 each operating at a symbol rate of 1/P times that of input signal Vin(t) with a timing offset of one symbol time of input signal Vin(t) relative to the neighboring slice. The collection of P slices 310 thus processes input signal Vin(t) in parallel to relax timing constraints on the individual slices. The number P of slices is sixty-four in this example. A single slice 310 producing a current two-bit data symbol HDQ(n) is illustrated and detailed for brevity. Adjacent slices like the depicted slice 310 resolve previous and subsequent symbols that can be used to compensate for precursor and postcursor ISI. ADC 305 can likewise be divided into slices to ease timing constraints.

Slice

310 includes a feed-forward stage 315 coupled to input node Din(n) to receive the six-bit digital values thereon. Stage 315 is an eleven-tap equalizer with three pre-cursor taps, a main tap, and seven post-cursor taps in this example. The three pre-cursor taps receive resolved data signals HDQ(n−3:n−1) from the three temporally adjacent slices 310 charged with interpreting the three preceding symbols, whereas the seven post-cursor taps receive resolved data signals HDQ(n+1:n+7) from the seven temporally adjacent slices 310 charged with interpreting the seven following symbols. Stage 315 cancels relatively low-frequency ISI, an effect referred to as “channel shortening.” Canceling relatively low-frequency ISI leaves a nine-bit equalized signal EQ1′ as a measure of relatively high-frequency ISI. Following the above-referenced Q notation, signal EQ1 has nine bits Q5.3, an implied sign bit, five integer bits, and three fractional bits. Delay elements 320 match the timing of equalized signal EQ1 to other equalization stages as detailed below.

Slice

310 includes another feed-forward stage 325 for equalizing input signal Din(n). Stage 325 is a seven-tap equalizer with three pre-cursor taps, a main tap, and three post-cursor taps in this embodiment. FFE stage 325 cancels considerable ISI to produce an equalized signal EQ2″. For short channels, equalized signal EQ2″ may be adequate for a hard sampler 330 to provide a two-bit, hard-decision data signal HDQ2″ with a suitably low BER. In that case a register 335 can be programmed to control (1) a multiplexer 340 to pass signal HDQ2″ as output data HDQ(n) and (2) a power controller 345 to power down components of receiver 300 that are not used to equalize input signal Din(n).

Slice

310 includes an ISI estimator 350 that can refine equalized signal EQ2″ from stage 325 using the measures of high-frequency ISI on offer as signal EQ1′ from equalizer FFE 315 and delay elements 320. ISI estimator 350 is divided into a cascade of R stages Stage1-StageR that provide progressively improved equalization at the cost of progressively increased power consumption. Considering the first stage Stage1, a decision circuit 355 resolves each nine-bit symbol of signal EQ2″ into a soft decision SDQ1″ placing the symbol into one of seven decision regions, one region for each of the four high-confidence interpretations of the incoming symbol (i.e., binary 00, 01, 10, and 11) and three low-confidence regions that represent uncertain ranges separating the high-confidence regions.

FIG. 4 depicts an eye diagram 400 of a PAM4 signal input to decision circuit 355. Diagram 400 is used to illustrate the distinction between high-confidence regions HC[XX] and low-confidence regions LC[YY]. The signal includes three

eyes

405, 410, and 415 separating the four voltage ranges that represent the four possible PAM4 symbol values at a sample instant around time zero. Sample values of signal EQ2″ that fall within the

ranges encompassing eyes

405, 410, and 415 do not clearly resolve the symbol under consideration. Values that fall within low-confidence range LC01 might represent either incoming symbol 10 or 01, for example. Decision circuit 355 slices incoming symbols relative to six threshold levels TH1 ₅:TH1 ₀to place each symbol into one of the seven low-confidence and high-confidence regions. Decision circuit 355 outputs soft-decision signal SDQ1″ expressing each sampled symbol as one of seven decisions d0-d6. An input/output diagram 450 below eye diagram 400 illustrates the operation of decision circuit 355 in one embodiment.

Returning to FIG. 3, a feed-forward equalizer 365 within ISI estimate stage Stage1 equalizes soft-decision signal SDQ1″ responsive to one precursor data symbol HDQ(n−1) and three postcursor data symbols HDQ(n+1:n+3). Stage Stage1 has no main tap in this example, but more, fewer, or different taps can be used here or elsewhere in other embodiments. An arithmetic operator 370 adds the resulting ISI estimate to signal EQ1′ from the first delay element 320 to produce an equalized signal EQ3-1 with a lower level of ISI than on signal EQ2″.

High confidence decisions are treated to tap values calibrated to fully offset ISI associated with those decisions. Low confidence decisions are treated to scaled tap values that reduce the effects of equalization and thus limit error propagation. The following Table 1 provides suitable tap values for each data level d0-d6, where a tap value of +h fully offsets ISI from a neighboring symbol of an amplitude representing a binary 11.

TABLE 1

Data Level	Confidence	Tap Value

d6	High	+h
d5	Low	+2h/3
d4	High	+h/3
d3	Low		0
d2	High	−h/3
d1	Low	−2h/3
d0	High	−h

If equalized signal EQ3-1 produces a sufficiently low BER then register 335 can be programmed to instruct multiplexer 340 to pass signal HDQ3-1, the output of another hard sampler 375, as output data HDQ(n). In the alternative, one or more additional ISI-estimation stages can be employed to further refine the signal. Because the equalized signal becomes progressively better as it moves through each stage of ISI estimator 350, the decision circuits 355 in each successive stage apply successively wider low-confidence regions. With reference to the last stage StageR, for example, decision circuit 355 slices an incoming equalized signal relative to six voltage levels THR₅:THR₀. The low-confidence regions are greater than those illustrated in FIG. 3 to reflect wider eyes and a reduced BER. As before, power controller 345 can power down circuitry not used to equalize and sample signal Vin(t).

FIG. 5 depicts ISI estimating Stage1 of FIG. 3 with decision threshold and equalizer tap adaptation circuitry 500. Threshold and tap-value selection and adjustment are well known to those of skill in the art so a detailed treatment of circuitry 500 is omitted.

ISIC stage Stage1 includes a series of delay elements T that each impose one symbol-time of delay. These elements provide four differently timed symbols from equalized signal EQ2, one symbol EQ2(n−1) in service of a precursor tap and three symbols EQ2(n+1:n+3) in service of three postcursor taps. Decisions from decision circuit 355 are conveyed to equalizer 365, which applies a respective set of tap weights to each of the precursor and postcursor slices. Arithmetic operator 370 combines the equalized outputs from equalizer 365 with equalized signal EQ1 to produce equalized signal EQ3-1.

Decision circuit

355 and equalizer 365 are divided into four taps in this example, one for each precursor and postcursor decision. To this end, decision circuit 355 includes four slicers 505(n−1;n+1:n+3) and equalizer 365 four multiplexers 510(n−1;n+1:n+3). Taking the precursor tap, for example, slicer 505(n−1) produces first-precursor decisions SDQ1(n−1). Multiplexer 510(n−1) applies the appropriate tap weight to each decision and passes on the resultant value to operator 370. The remaining filter taps work the same way except that the applied tap weights are tailored for the impact of the aggressor symbol they are calibrated to counteract.

While the present invention has been described in connection with specific embodiments, variations of these embodiments will be obvious to those of ordinary skill in the art. For example, embodiments of the invention may be adapted for use with multi-PAM signals beyond PAM-4. Moreover, some components are shown directly connected to one another while others are shown connected via intermediate components. In each instance the method of interconnection, or “coupling,” establishes some desired electrical communication between two or more circuit nodes, terminals, or ports. Such coupling may often be accomplished using various circuit configurations, as will be understood by those of skill in the art. Therefore, the spirit and scope of the appended claims should not be limited to the foregoing description. Where U.S. law applies, only those claims specifically reciting “means for” or “step for” should be construed in the manner required under the sixth paragraph of 35 U.S.C. § 112.

Claims

What is claimed is:

1. A method of interpreting an input signal expressing a series of symbols, each of the symbols representing one of a set of possible symbol values, the method comprising:

equalizing the input signal to produce a first equalized version of the input signal;

slicing the first equalized version of the input signal to obtain high-confidence decisions and low-confidence decisions;

equalizing the high-confidence decisions and the low-confidence decisions to produce a second equalized version of the input signal; and

combining the first equalized version of the input signal with the second equalized version of the input signal to produce a third equalized version of the input signal.

2. The method of claim 1, further comprising sampling the third equalized version of the input signal to recover the series of symbols.

3. The method of claim 1, wherein equalizing the input signal to produce the first equalized version of the input signal applies a first set of tap values and equalizing the high-confidence decisions and the low-confidence decisions applies a second set of tap values different from the first set of tap values.

4. The method of claim 3, wherein the first set of tap values includes a main-tap value and the second set of tap values omits the main-tap value.

5. The method of claim 1, further comprising:

sampling the third equalized version of the input signal to obtain second high-confidence decisions and second low-confidence decisions;

equalizing the second high-confidence decisions and the second low-confidence decisions to produce a fourth equalized version of the input signal; and

combining the first equalized version of the input signal with the fourth equalized version of the input signal to produce a fifth equalized version of the input signal.

6. The method of claim 5, further comprising sampling the fifth equalized version of the input signal to recover the series of symbols.

7. A method of interpreting an input signal expressing a series of symbols, each of the symbols representing one of a set of possible symbol values, the method comprising:

equalizing the input signal to produce a first equalized version of the input signal and a second equalized version of the input signal;

sampling the second equalized version of the input signal to obtain high-confidence decisions and low-confidence decisions;

equalizing the high-confidence decisions and the low-confidence decisions to produce a third equalized version of the input signal; and

combining the first equalized version of the input signal with the third equalized version of the input signal to produce a fourth equalized version of the input signal.

8. The method of claim 7, further comprising sampling the fourth equalized version of the input signal to recover the series of symbols.

9. The method of claim 7, wherein equalizing the input signal to produce the first equalized version of the input signal applies a first set of tap values and equalizing the high-confidence decisions and the low-confidence decisions applies a second set of tap values different from the first set of tap values.

10. The method of claim 9, wherein the first set of tap values includes a main-tap value and the second set of tap values omits the main-tap value.

11. The method of claim 7, further comprising:

sampling the fourth equalized version of the input signal to obtain second high-confidence decisions and second low-confidence decisions;

equalizing the second high-confidence decisions and the second low-confidence decisions to produce a fifth equalized version of the input signal; and

combining the first equalized version of the input signal with the fifth equalized version of the input signal to produce a sixth equalized version of the input signal.

12. The method of claim 11, further comprising sampling the sixth equalized version of the input signal to recover the series of symbols.

13. A feed-forward equalizer comprising:

a first node to receive a first equalized version of an input signal;

a second node to receive a second equalized version of the input signal;

a decision circuit coupled to the second node to obtain high-confidence decisions and low-confidence decisions of the second equalized version of the input signal;

at least one equalization tap coupled to the decision circuit to equalize the high-confidence decisions and the low-confidence decisions to produce an interference estimate of the input signal; and

an arithmetic operator coupled to the first node and the equalization tap, the arithmetic operator to combine the first equalized version of the input signal with the interference estimate of the input signal into a third equalized version of the input signal.

14. The feed-forward equalizer of claim 13, further comprising a hard sampler coupled to the arithmetic operator to produce hard samples of the third equalized version of the input signal.

15. The feed-forward equalizer of claim 13, further comprising a second decision circuit coupled to the arithmetic operator to obtain high-confidence decisions and low-confidence decisions of the third equalized version of the input signal.

16. The feed-forward equalizer of claim 15, further comprising:

a second equalization tap coupled to the second decision circuit to equalize the high-confidence decisions and the low-confidence decisions of the third equalized version of the input signal to produce a second interference estimate of the input signal, and

a second arithmetic operator to combine the first equalized version of the input signal with the second interference estimate of the input signal to produce a fourth equalized version of the input signal.

17. The feed-forward equalizer of claim 16, further comprising a hard sampler coupled to the second arithmetic operator to produce hard samples of the fourth equalized version of the input signal.

18. The feed-forward equalizer of claim 17, further comprising a second hard sampler coupled to the first-mentioned arithmetic operator to produce second hard samples of the fourth equalized version of the input signal.

19. The feed-forward equalizer of claim 18, further comprising a multiplexer to select between the hard samples of the combined first and fourth equalized versions of the input signal and the second hard samples of the combined first and second equalized versions of the input signal.

20. The feed-forward equalizer of claim 13, the at least one equalization tap consisting of equalization taps from a group of precursor and postcursor taps, the feed-forward equalizer further comprising a feed-forward stage to produce the first equalized version of the input signal from the input signal, the feed-forward stage including a main tap.